Gork, the hero we need

the street finds its own uses for things

Over several posts I want to explore the idea that the adoption of epistemically sound, trustworthy AI advisors is a good path forward for maximizing the benefits and minimizing the risks of AI.

—

X.com is not a place where one often encounters the better angels of our nature. The culture of the platform and the incentives for controversy push towards a PvP, adversarial, truth-is-secondary-to-a-good-dunk worldview.

But, even in the most wretched wastes, a hero might emerge:

Gork is a misspelling of Grok, x.ai’s AI. It became a meme in around April 2025 as users were starting to frequently tag in Grok to fact check claims made on the platform, and at times misspelling it as Gork. I’m fond of Gork - in my head canon he’s a folk hero. Gork never dallied with nazism; Gork just wants to help the common man know ‘is this true’. But, for the sake of clarity, I’ll refer to its institutional name Grok.

I tried to look into how often Grok is used for fact checking - the data is sparse. An Aljazeera article claims that there were 2.3 million total calls in one week in June 5 and June 12, and that in two of the cases the ratio was ~68% of the calls asking for image authentication, and in another 20% asking for a fact check (70% if you include asks for fact check or additional context.)

If we use the Aljazeera claims as a lower and upper bound, this would be 460,000 to 1,610,000 fact checks a week. Personally I’m inclined to take a low estimate, and suggest that it’s somewhere between 500,000 and 1,000,000 fact checks¹.

Tagging in Grok for a public fact check is a grassroots driven enactment of a novel AI social process to adjudicate and investigate factual claims. It’s at times adversarial - many people are using this to privilege a hypothesis that the person they are fact checking is wrong² - but Grok has this power in the arena because AIs are widely trusted as sources of information.

The Collective Intelligence Project has been running global dialogues on people’s relationships to AI; over their last three dialogues from March to August 2025 AI chatbots have ranked as the most trusted ‘group’, only falling behind research institutions and family doctors.

This is wild. A member of the public trusts an AI to act in their best interest more than companies, more than elected officials, more than their spiritual leaders!

You can read this as either a stunning indictment of our existing institutions (which, fair), as a symptom of a public deluded by AI sycophancy, or as a point of excitement that post the revolt-of-the-public, when trust in institutions has declined continuously for decades, there are new “institutions” that people actually trust.

Personally I choose excitement. Novel institutions that people trust enough to adjudicate difficult-to-verify fact claims are rare. I think many of our most difficult societal problems come down to competing factual claims and values that nobody can orient and reason about alone - this directly helps with factual claims, and I suspect can act as a bridging system for value disagreement.

Importantly, I am not claiming that Grok or any of the existing LLM systems as they stand today are adequate, truth seeking systems. They are not robust enough, they do not handle competing ontologies well, they can not confidently extend causal inferences and counterfactual reasoning well, they can be manipulated (there’s a troublesome intern at Grok that’s done this from time to time).

But I think we could be on a tech path where we get those improvements through a combination of scale and external scaffolded knowledge systems and tools that the AIs can use, and which the public can trust.

I’ll address the roadmap in later posts, but a key implication is that this emergent trust is a very rare and fragile asset. Its value depends on the perceived neutrality and competency of the AI systems. If we hope these systems can advise the public and decision makers on high-stakes problems, we should avoid attempts to censor or neuter their outputs, which would likely destroy the trust that makes them useful in the first place.

@gork that’s true, right?

1. It’s a drop in the bucket of total tweets per week (roughly 3.5 billion), but then again the vast majority of x posts are not making factual claims.

2. Attempts to invoke Grok in this way can go awry